Marcin Miłkowski

I argue that there are no plausible non-representational explanations of episodes of hallucination. To make the discussion more specific, I focus on visual hallucinations in Charles Bonnet syndrome. I claim that the character of such hallucinatory experiences cannot be explained away non-representationally, for they cannot be taken as simple failures of cognizing or as failures of contact with external reality—such failures being the only genuinely non-representational explanations of hallucinations and cognitive errors in general. I briefly introduce a recent computational model of hallucination, which relies on generative models in the brain, and argue that the model is a prime example of a representational explanation referring to representational mechanisms. The notion of the representational mechanism is elucidated, and it is argued that hallucinations—and other kinds of representations—cannot be exorcised from the cognitive sciences.

This paper examines the interplay between integrative explanatory pluralism and the quest for unified theories. We argue that when grounded in virtues associated with satisfactory explanations, integrative pluralism exhibits an inherent instability stemming from the conflict between the demand for unity and the commitment to preserving a patchwork of disparate partial explanations. A case study in cognitive science illuminates the challenges of maintaining both systematicity and depth in explanations within this framework. While this instability does not render integrative pluralism fundamentally flawed, it stresses the importance of a diachronic analysis of scientific dynamics and norms. The conclusion highlights the continued value of integrative pluralism in interdisciplinary research programs, while emphasizing its role as a temporary rather than permanent approach.

This chapter argues that the extended mind approach to cognition can be distinguished from its alternatives, such as embedded cognition and distributed cognition, not only in terms of metaphysics, but also in terms of epistemology. In other words, it cannot be understood in terms of a mere verbal redefinition of cognitive processing. This is because the extended mind approach differs in its theoretical virtues compared to competing approaches to cognition. The extended mind approach is thus evaluated in terms of its theoretical virtues, both essential to empirical adequacy and those that are ideal desiderata for scientific theories. While the extended mind approach may have similar internal consistency and empirical adequacy compared to other approaches, it may be more problematic in terms of its generality and simplicity as well as unificatory properties due to the cognitive bloat and the motley crew objections.

This paper presents an argument for the realism about mechanisms, contents, and vehicles of mental representation at both the personal and subpersonal levels, and showcases its role in instrumental rationality and proper cognitive functioning. By demonstrating how misrepresentation is necessary for learning from mistakes and explaining certain failures of action, we argue that fallible rational agents must have mental representations with causally relevant vehicles of content. Our argument contributes to ongoing discussions in philosophy of mind and cognitive science by challenging anti-realist views about the nature of mental representation, and by highlighting the importance of understanding how different agents can misrepresent in pursuit of their goals. While there are potential rebuttals to our claim, our opponents must explain how agents can be rational without having mental representations. This is because mental representation is grounded in rationality.

In light of the recent credibility crisis in psychology, this paper argues for a greater emphasis on theorizing in scientific research. Although reliable experimental evidence, preregistration, methodological rigor, and new computational frameworks for modeling are important, scientific progress also relies on properly functioning theories. However, the current understanding of the role of theorizing in psychology is lacking, which may lead to future crises. Theories should not be viewed as mere speculations or simple inductive generalizations. To address this issue, the author introduces a framework called “cognitive metascience,” which studies the processes and results of evaluating scientific practice. This study should proceed both qualitatively, as in traditional science and technology studies and cognitive science, and quantitatively, by analyzing scientific discourse using language technology. By analyzing theories as cognitive artifacts that support cognitive tasks, this paper aims to shed more light on their nature. This perspective reveals that multiple distinct theories serve entirely different roles, and studying these roles, along with their epistemic vices and virtues, can provide insight into how theorizing should proceed. The author urges a change in research culture to appreciate the variety of distinct theories and to systematically advance scientific progress. 

Levenstein et al. aptly highlight some of the foundational issues in theoretical neuroscience, such as the role of abstraction and idealization in providing scientific explanations and understanding, and distinguishing under which conditions neuroscientific models provide genuine explanations, or mere descriptions, predictions, or control. The authors rightly emphasize that philosophers of science can also gain valuable insights from the vast body of neuroscience literature, by employing methods of digital humanities, such as text mining, in line with the cognitive metascience approach (Miłkowski, 2023). Conversely, neuroscientists can benefit from the latest research in philosophy of science, particularly on the epistemic norms of diverse kinds of explanations in neuroscience, e.g., topological (Kostić, 2020), computational (Miłkowski, 2013; Chirimuuta, 2018), dynamical (Favela, 2020), and functional (Egan, 2017). Understanding the diverse goals of theories is conducive to fruitful exploratory and explanatory scientific practice. Some forms of scientific understanding are not always grounded in explanations, but also in diverse theoretical representations, e.g., diagrams, taxonomies, and diagnostic manuals (Miłkowski, 2023). Philosophical theories of understanding can also integrate different kinds of explanations in a single framework (Khalifa et al., 2022). The need for these interdisciplinary connections is particularly vivid in the case of network models that are increasingly being used in neuroscience (Kostić et al., 2020), and especially when it comes to evaluating their explanatory power (Kostić, 2020; Kostić and Khalifa, 2023). In conclusion, we strongly support Levenstein et al.'s call for greater collaboration between philosophers of science and neuroscientists on the foundational issues in neuroscience. Embracing this interdisciplinary approach is essential for advancing our knowledge of the brain and its function.

Three special issues of Entropy journal have been dedicated to the topics of “InformationProcessing and Embodied, Embedded, Enactive Cognition”. They addressed morphological computing, cognitive agency, and the evolution of cognition. The contributions show the diversity of views present in the research community on the topic of computation and its relation to cognition. This paper is an attempt to elucidate current debates on computation that are central to cognitive science. It is written in the form of a dialog between two authors representing two opposed positions regarding the issue of what computation is and could be, and how it can be related to cognition. Given the different backgrounds of the two researchers, which span physics, philosophy of computing and information, cognitive science, and philosophy, we found the discussions in the form of Socratic dialogue appropriate for this multidisciplinary/cross-disciplinary conceptual analysis. We proceed as follows. First, the proponent (GDC) introduces the info-computational framework as a naturalistic model of embodied, embedded, and enacted cognition. Next, objections are raised by the critic (MM) from the point of view of the new mechanistic approach to explanation. Subsequently, the proponent and the critic provide their replies. The conclusion is that there is a fundamental role for computation, understood as information processing, in the understanding of embodied cognition.

One of the critical issues in the philosophy of science is to understand scientific knowledge. This paper proposes a novel approach to the study of reflection on science, called “cognitive metascience”. In particular, it offers a new understanding of scientific knowledge as constituted by various kinds of scientific representations, framed as cognitive artifacts. It introduces a novel functional taxonomy of cognitive artifacts prevalent in scientific practice, covering a huge diversity of their formats, vehicles, and functions. As a consequence, toolboxes, conceptual frameworks, theories, models, and individual hypotheses can be understood as artifacts supporting our cognitive performance. It is also shown that by empirically studying how artifacts function, we may discover hitherto undiscussed virtues and vices of these scientific representations. This paper relies on the use of language technology to analyze scientific discourse empirically, which allows us to uncover the metascientific views of researchers. This, in turn, can become part of normative considerations concerning virtues and vices of cognitive artifacts.

In most accounts of realization of computational processes by physical mechanisms, it is presupposed that there is one-to-one correspondence between the causally active states of the physical process and the states of the computation. Yet such proposals either stipulate that only one model of computation is implemented, or they do not reflect upon the variety of models that could be implemented physically. In this paper, I claim that mechanistic accounts of computation should allow for a broad variation of models of computation. In particular, some non-standard models should not be excluded a priori. The relationship between mathematical models of computation and mechanistically adequate models is studied in more detail.

The article presents the interdisciplinary approach of Edwin Hutchins, analyzing his conception of distributed cognition as probably the most important and lasting contribution of anthropology to the repertoire of theoretical tools in cognitive science. At the same time, this conception resulted in one of the most interesting relationships between cognitive science and social sciences. These relationships are made possible by the assumptions of Hutchins’ conception, which directly contribute to interdisciplinary collaboration. His account of distributed cognition has enormous potential, allowing the integration of research into cognitive and social processes. This is also because it breaks with methodological individualism.

The predictive processing account of action, cognition, and perception is one of the most influential approaches to unifying research in cognitive science. However, its promises of grand unification will remain unfulfilled unless the account becomes theoretically robust. In this paper, we focus on empirical commitments of PP, since they are necessary both for its theoretical status to be established and for explanations of individual phenomena to be falsifiable. First, we argue that PP is a varied research tradition, which may employ various kinds of scientific representations, differing in the scope of empirical commitments they entail. Two major perspectives on PP qua cognitive theory may then be distinguished: generalized vs. hierarchical. The first one fails to provide empirical detail, and the latter constrains possible physical implementations. However, we show that even hierarchical PP is insufficiently restrictive to disallow incorrect models and may be adjusted to explain any neurocognitive phenomenon–including non-existent or impossible ones–through flexible adjustments. This renders PP a universal modeling tool with an unrestricted number of degrees of freedom. Therefore, in contrast with declarations of its proponents, it should not be understood as a unifying theoretical perspective, but as a computational framework, possibly informing further theory development in cognitive science.

Alan Turing’s influence on subsequent research in artificial intelligence is undeniable. His proposed test for intelligence remains influential. In this paper, I propose to analyze his conception of intelligence by relying on traditional close reading and language technology. The Turing test is interpreted as an instance of conceptual engineering that rejects the role of the previous linguistic usage, but appeals to intuition pumps instead. Even though many conceive his proposal as a prime case of operationalism, it is more plausibly viewed as a stepping stone toward a future theoretical construal of intelligence in mechanical terms. To complete this picture, his own conceptual network is analyzed through the lens of distributional semantics over the corpus of his written work. As it turns out, Turing’s conceptual engineering of the notion of intelligence is indeed quite similar to providing a precising definition with the aim of revising the usage of the concept. However, that is not its ultimate aim: Turing is after a rich theoretical understanding of thinking in mechanical, i.e., computational, terms.

Is the mathematical function being computed by a given physical system determined by the system’s dynamics? This question is at the heart of the indeterminacy of computation phenomenon (Fresco et al. [unpublished]). A paradigmatic example is a conventional electrical AND-gate that is often said to compute conjunction, but it can just as well be used to compute disjunction. Despite the pervasiveness of this phenomenon in physical computational systems, it has been discussed in the philosophical literature only indirectly, mostly with reference to the debate over realism about physical computation and computationalism. A welcome exception is Dewhurst’s ([2018]) recent analysis of computational individuation under the mechanistic framework. He rejects the idea of appealing to semantic properties for determining the computational identity of a physical system. But Dewhurst seems to be too quick to pay the price of giving up the notion of computational equivalence. We aim to show that the mechanist need not pay this price. The mechanistic framework can, in principle, preserve the idea of computational equivalence even between two different enough kinds of physical systems, say, electrical and hydraulic ones.

A novel account of semantic information is proposed. The gist is that structural correspondence, analysed in terms of similarity, underlies an important kind of semantic information. In contrast to extant accounts of semantic information, it does not rely on correlation, covariation, causation, natural laws, or logical inference. Instead, it relies on structural similarity, defined in terms of correspondence between classifications of tokens into types. This account elucidates many existing uses of the notion of information, for example, in the context of scientific models and structural representations in cognitive science. It is poised to open a new research program concerned with various kinds of semantic information, its functions, and its measurement.

In his recent book, Daniel Dennett defends a novel account of semantic information in terms of design worth getting (Dennett, 2017). While this is an interesting proposal in itself, my purpose in this commentary is to challenge several of Dennett’s claims. First, he argues that semantic information can be transferred without encoding and storing it. Second, this lack of encoding is what makes semantic information unmeasurable. However, the argument for both these claims, presented by Dennett as an intuition pump, is invalid.

The debate between the defenders of explanatory unification and explanatory pluralism has been ongoing from the beginning of cognitive science and is one of the central themes of its philosophy. Does cognitive science need a grand unifying theory? Should explanatory pluralism be embraced instead? Or maybe local integrative efforts are needed? What are the advantages of explanatory unification as compared to the benefits of explanatory pluralism? These questions, among others, are addressed in this Synthese’s special issue. In the introductory paper, we discuss the background of the questions, distinguishing integrative theorizing from building unified theories. On the one hand, integrative efforts involve collaboration between various disciplines, fields, approaches, or theories. These efforts could even be quite temporary, without establishing any long-term institutionalized fields or disciplines, but could also contribute to developing new interfield theories. On the other hand, unification can rely on developing complete theories of mechanisms and representations underlying all cognition, as Newell’s “unified theories of cognition”, or may appeal to grand principles, as predictive coding. Here, we also show that unification in contemporary cognitive science goes beyond reductive unity, and may involve various forms of joint efforts and division of explanatory labor. This conclusion is one of the themes present in the content of contributions constituting the special issue.

Predictive processing (PP) has been repeatedly presented as a unificatory account of perception, action, and cognition. In this paper, we argue that this is premature: As a unifying theory, PP fails to deliver general, simple, homogeneous, and systematic explanations. By examining its current trajectory of development, we conclude that PP remains only loosely connected both to its computational framework and to its hypothetical biological underpinnings, which makes its fundamentals unclear. Instead of offering explanations that refer to the same set of principles, we observe systematic equivocations in PP‐based models, or outright contradictions with its avowed principles. To make matters worse, PP‐based models are seldom empirically validated, and they are frequently offered as mere just‐so stories. The large number of PP‐based models is thus not evidence of theoretical progress in unifying perception, action, and cognition. On the contrary, we maintain that the gap between theory and its biological and computational bases contributes to the arrested development of PP as a unificatory theory. Thus, we urge the defenders of PP to focus on its critical problems instead of offering mere re‐descriptions of known phenomena, and to validate their models against possible alternative explanations that stem from different theoretical assumptions. Otherwise, PP will ultimately fail as a unified theory of cognition.

I argue that there are no plausible non-representational explanations of episodes of hallucination. To make the discussion more specific, I focus on visual hallucinations in Charles Bonnet syndrome. I claim that the character of such hallucinatory experiences cannot be explained away non-representationally, for they cannot be taken as simple failures of cognizing or as failures of contact with external reality—such failures being the only genuinely non-representational explanations of hallucinations and cognitive errors in general. I briefly introduce a recent computational model of hallucination, which relies on generative models in the brain, and argue that the model is a prime example of a representational explanation referring to representational mechanisms. The notion of the representational mechanism is elucidated, and it is argued that hallucinations—and other kinds of representations—cannot be exorcised from the cognitive sciences.

Is there a field of social intelligence? Many various disciplines approach the subject and it may only seem natural to suppose that different fields of study aim at explaining different phenomena; in other words, there is no special field of study of social intelligence. In this paper, I argue for an opposite claim. Namely, there is a way to integrate research on social intelligence, as long as one accepts the mechanistic account to explanation. Mechanistic integration of different explanations, however, comes at a cost: mechanism requires explanatory models to be fairly complete and realistic, and this does not seem to be the case for many models concerning social intelligence, especially models of economical behavior. Such models need either be made more realistic, or they would not count as contributing to the same field. I stress that the focus on integration does not lead to ruthless reductionism; on the contrary, mechanistic explanations are best understood as explanatorily pluralistic.

In this paper, we defend a novel, multidimensional account of representational unification, which we distinguish from integration. The dimensions of unity are simplicity, generality and scope, non-monstrosity, and systematization. In our account, unification is a graded property. The account is used to investigate the issue of how research traditions contribute to representational unification, focusing on embodied cognition in cognitive science. Embodied cognition contributes to unification even if it fails to offer a grand unification of cognitive science. The study of this failure shows that unification, contrary to what defenders of mechanistic explanation claim, is an important mechanistic virtue of research traditions.

The focus of this special issue of Theory & Psychology is on explanatory mechanisms in psychology, especially on problems of particular prominence for psychological science such as theoretical integration and unification. Proponents of the framework of mechanistic explanation claim, in short, that satisfactory explanations in psychology and related fields are causal. They stress the importance of explaining phenomena by describing mechanisms that are responsible for them, in particular by elucidating how the organization of component parts and operations in mechanisms gives rise to phenomena in certain conditions. We hope for cross-pollination between philosophical approaches to explanation and experimental psychology, which could offer methodological guidance, in particular where mechanism discovery and theoretical integration are at issue. Contributions in this issue pertain to theoretical integration and unification of psychology as well as the growing importance of causal mechanistic explanations in psychological science.

In this paper, I argue that embodied cognition, like many other research traditions in cognitive science, offers mostly fallible research heuristics rather than grand principles true of all cognitive processing. To illustrate this claim, I discuss Aizawa’s rebuttal of embodied and enactive accounts of vision. While Aizawa’s argument is sound against a strong reading of the enactive account, it does not undermine the way embodied cognition proceeds, because the claim he attacks is one of fallible heuristics. These heuristics may be helpful in developing models of cognition in an interdisciplinary fashion. I briefly discuss the issue of whether this fallibility actually makes embodied cognition vulnerable to charges of being untestable or non-scientific. I also stress that the historical approach to this research tradition suggests that embodied cognition is not poised to become a grand unified theory of cognition.

In this paper, we focus on the development of geometric cognition. We argue that to understand how geometric cognition has been constituted, one must appreciate not only individual cognitive factors, such as phylogenetically ancient and ontogenetically early core cognitive systems, but also the social history of the spread and use of cognitive artifacts. In particular, we show that the development of Greek mathematics, enshrined in Euclid’s Elements, was driven by the use of two tightly intertwined cognitive artifacts: the use of lettered diagrams; and the creation of linguistic formulae. Together, these artifacts formed the professional language of geometry. In this respect, the case of Greek geometry clearly shows that explanations of geometric reasoning have to go beyond the confines of methodological individualism to account for how the distributed practice of artifact use has stabilized over time. This practice, as we suggest, has also contributed heavily to the understanding of what mathematical proof is; classically, it has been assumed that proofs are not merely deductively correct but also remain invariant over various individuals sharing the same cognitive practice. Cognitive artifacts in Greek geometry constrained the repertoire of admissible inferential operations, which made these proofs inter-subjectively testable and compelling. By focusing on the cognitive operations on artifacts, we also stress that mental mechanisms that contribute to these operations are still poorly understood, in contrast to those mechanisms which drive symbolic logical inference.

Is the mathematical function being computed by a given physical system determined by the system’s dynamics? This question is at the heart of the indeterminacy of computation phenomenon (Fresco et al. [unpublished]). A paradigmatic example is a conventional electrical AND-gate that is often said to compute conjunction, but it can just as well be used to compute disjunction. Despite the pervasiveness of this phenomenon in physical computational systems, it has been discussed in the philosophical literature only indirectly, mostly with reference to the debate over realism about physical computation and computationalism. A welcome exception is Dewhurst’s ([2018]) recent analysis of computational individuation under the mechanistic framework. He rejects the idea of appealing to semantic properties for determining the computational identity of a physical system. But Dewhurst seems to be too quick to pay the price of giving up the notion of computational equivalence. We aim to show that the mechanist need not pay this price. The mechanistic framework can, in principle, preserve the idea of computational equivalence even between two different enough kinds of physical systems, say, electrical and hydraulic ones.

The purpose of this paper is to argue against the claim that morphological computation is substantially different from other kinds of physical computation. I show that some (but not all) purported cases of morphological computation do not count as specifically computational, and that those that do are solely physical computational systems. These latter cases are not, however, specific enough: all computational systems, not only morphological ones, may (and sometimes should) be studied in various ways, including their energy efficiency, cost, reliability, and durability. Second, I critically analyze the notion of “offloading” computation to the morphology of an agent or robot, by showing that, literally, computation is sometimes not offloaded but simply avoided. Third, I point out that while the morphology of any agent is indicative of the environment that it is adapted to, or informative about that environment, it does not follow that every agent has access to its morphology as the model of its environment.

Replicability and reproducibility of computational models has been somewhat understudied by “the replication movement.” In this paper, we draw on methodological studies into the replicability of psychological experiments and on the mechanistic account of explanation to analyze the functions of model replications and model reproductions in computational neuroscience. We contend that model replicability, or independent researchers' ability to obtain the same output using original code and data, and model reproducibility, or independent researchers' ability to recreate a model without original code, serve different functions and fail for different reasons. This means that measures designed to improve model replicability may not enhance (and, in some cases, may actually damage) model reproducibility. We claim that although both are undesirable, low model reproducibility poses more of a threat to long-term scientific progress than low model replicability. In our opinion, low model reproducibility stems mostly from authors' omitting to provide crucial information in scientific papers and we stress that sharing all computer code and data is not a solution. Reports of computational studies should remain selective and include all and only relevant bits of code.